Weighted Voting Systems: Reliability Versus Rapidity: Gregory Levitin

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

Reliability Engineering and System Safety 89 (2005) 177–184

www.elsevier.com/locate/ress

Weighted voting systems: reliability versus rapidity


Gregory Levitin*
Reliability Department, Planning, Development and Technology Division, Israel Electric Corporation Ltd, P.O. Box 10, Haifa 31000, Israel
Received 25 February 2004; accepted 17 August 2004
Available online 8 October 2004

Abstract

The weighted voting system (WVS) consists of n units that each provide a binary decision (0 or 1) or abstain from voting. Each
unit has its own individual weight. System output is 1 if the cumulative weight of all 1-opting units is at least a pre-specified fraction
t of the cumulative weight of all non-abstaining units. Otherwise, system output is 0. The system input is either 0 or 1. Every unit is
characterized by probability of making decisions 0 and 1 and by probability of abstaining for each input. The system fails if its output
is not equal to its input.
This paper shows that if the WVS consists of units that need different time to produce their outputs, the decision time of the entire system
depends on the distribution of unit weights and on the value of t. It shows also that a tradeoff exists between the system reliability and its
rapidity.
An algorithm that finds the system parameters maximizing its reliability under constraint imposed on the expected system decision time is
suggested. Illustrative examples are presented.
q 2004 Elsevier Ltd. All rights reserved.

1. Introduction implicit form. Therefore the units are subject to the


following three errors:
Voting systems (k-out-of-n systems with multiple failure
modes) are widely used in human organization systems as 1. Acceptance of a proposition that should be rejected (fault
well as in technical decision making systems. The reliability of being too optimistic),
of these systems has been studied in [1–6]. The weighted 2. Rejection of a proposition that should be accepted (fault
voting systems (WVS) are generalization of the voting of being too pessimistic),
systems. WVS are intensively studied in recent years 3. Abstaining from voting (fault of being unavailable or
[7–12]. The applications of WVS can be found in imprecise indecisive).
data handling [13], safety monitoring and self-testing [14],
This can be modeled by considering system input I being
multi-channel signal processing [15], pattern recognition
either 1 (proposition to be accepted) or 0 (proposition to be
and target detection [7], etc.
rejected) which is supplied to each unit. Each unit j produces
A WVS makes a decision about propositions based on
its decision (unit output) dj(I) which can be 1, 0 or x (in the
the decisions of n statistically independent individual units
case of abstention). Inequality dj(I)sI means that the
of which it consists (for example, in target detecting system
decision made by the unit is wrong. The above listed errors
speed detectors and heat radiation detectors provide the
can be expressed as
system with their individual decisions without communicat-
ing among themselves). Each proposition is a priori right or 1. dj(0)Z1 (unit fails stuck-at-1),
wrong but this information is available for the units in 2. dj(1)Z0 (unit fails stuck-at-0),
3. dj(I)Zx (unit fails stuck-at-x).

Accordingly, reliability of each unit j can be character-


* Tel.: C972 4 8183726; fax: C972 4 8183790.
E-mail address: levitin@iec.co.il
ized by probabilities of these errors: qðjÞ ðjÞ
01 for the first one, q10

0951-8320/$ - see front matter q 2004 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ress.2004.08.017
178 G. Levitin / Reliability Engineering and System Safety 89 (2005) 177–184

Nomenclature
t threshold factor, (0%t%1) tj time when the decision dj(I) of voting unit j
n number of units belonging to WVS becomes available
K number of different voting results in WVS T expected WVS decision time
I WVS input (proposition to be accepted or T* maximum allowed expected decision time
rejected), I2{0,1} Wm1 total weight of units that have decision time not
Pr{v} probability of event v greater than tm and vote for the proposition
pi Pr{IZi} acceptance
dj(I) decision (output) of individual unit j Wm0 total weight of units that have decision time not
D(I) decision of the entire WVS greater than tm and vote for the proposition
qðjÞ
is Pr{dj(I)ZsjIZi},i2{0,1},s2{0,1,x} rejection
Qis Pr{D(I)ZsjIZi}, i2{0,1},s2{0,1,x} hm
I0 probability that the proposition I is rejected at
R Pr{D(I)ZI}, WVS reliability time tm
UIm ðzÞ U-function representing all the possible different hm
I1 probability that the proposition I is accepted at
states of a subsystem containing m fastest voting time tm
units when system input is I 1(x) unity function: 1(TRUE)Z1, 1(FALSE)Z0
w0j ‘negative’ weight of unit j (weight of unit j when
Acronyms1
dj(I)Z0)
WVS weighted voting system
w1j ‘positive’ weight of unit j (weight of unit j when
UF U(z)-function
dj(I)Z1)

for the second one, qðjÞ ðjÞ


1x and q0x for the third one (note that the threshold can be chosen in such a way that maximizes
stuck-at-x probabilities can be different for inputs IZ0 and the entire WVS reliability:
IZ1).
In this paper we consider an asymmetric WVS (first ðw01 ; w11 ; .; w0n ; w1n ; tÞ Z argfRðw01 ; w11 ; .; w0n ; w1n ; tÞ/ maxg:
suggested in [9]) which is able to take advantage of (1)
knowledge about statistical asymmetry of voting units
The algorithm for WVS optimization according to this
(asymmetric probabilities of making correct decisions with formulation is presented in Ref. [9].
respect to the input I) and therefore has greater reliability The previous studies of WVS do not address the aspect of
than the symmetric WVS. In such system each voting unit j the system rapidity in making decisions. The first paper that
has two weights that express its relative importance in the considers the decision time factor is by Xie and Pham [11].
WVS: ‘negative’ weight w0j which is assigned to the unit In this paper, it is assumed that the probabilities of making
when it votes for the proposition rejection and ‘positive’ mistakes decreases with the time for each voting unit but the
weight w1j which is assigned to the unit when it votes for the cost of decision making by the entire system increases with
proposition acceptance. To make a decision about prop- the time. An algorithm for finding the optimal stopping time
osition acceptance, the system incorporates all the unit (the time when all the units are obliged to vote) is
decisions into a unanimous system output D (Fig. 1). The suggested. This model is relevant for human organizations
proposition is rejected by the WVS (D(I)Z0) if the total in which the requirement of a limited time to complete
weight of units voting for its acceptance is less than a pre- the report (to make the decision) affects the results of the
specified fraction t of total weight of not abstaining units experts’ evaluations together with their expertise.
(t is usually referred to as threshold factor). The WVS In many technical systems, the time when the output
abstains (D(I)Zx) if all of its voting units abstain. (decision) of each voting unit is available is predetermined.
The system fails if D(I)sI. The entire WVS reliability For example, the decision time of a chemical analyzer is
can be defined as RZPr{D(I)ZI}. One can see that the determined by the time of a chemical reaction. The decision
system reliability is a function of reliabilities of units it time of a target detection radar system is determined by the
consists of. The reliability characteristics of WVS units as time of the radio signal return and by the time of signal
well as the propositions probability distribution can be processing by the electronic subsystem. In both these cases,
elicited from historical statistics. In technical systems, the variation of the decision times is usually negligible small.
probabilities of different kinds of errors can be obtained for On the contrary, the decision time of the entire WVS
each unit with high precision by intensive testing. The entire composed from voting units with different constant decision
WVS reliability also depends on the unit weights and the
threshold. While the units’ reliabilities usually cannot be 1
The singular and plural forms of acronyms are always spelled the
changed when the WVS is built, the weights and same.
G. Levitin / Reliability Engineering and System Safety 89 (2005) 177–184 179

a pre-specified value T*:

ðw01 ; w11 ; .; w0n ; w1n ; tÞ


Z argfRðw01 ; w11 ; .; w0n ; w1n ; tÞ/ max (2)
jTðw01 ; w11 ; .; w0n ; w1n ; tÞ% Tg:
This paper presents an algorithm for solving this
optimization problem. Section 2 of the paper presents the
voting model. Section 3 describes an algorithm for
reliability and performance evaluation. Section 4 is
devoted to optimization technique.

Fig. 1. Example of asymmetric WVS with nZ6. 2. The weighted voting model

times can vary. As it was mentioned in [11], the system does Consider the WVS consisting of n voting units
not need to wait for decisions of slow voting units, as long, characterized by error probabilities qðjÞ ðjÞ ðjÞ ðjÞ
01 , q10 , q1x , q0x and
as the system can make a correct decision with reliability decision times tj. The WVS incorporates all the unit outputs
higher than a pre-specified level. Moreover, in some cases into a unanimous system output D using the following
the decisions of the slow voting units do not affect the threshold based decision rule:
8 X X X
>
> 1; if w1j dj ðIÞR t ½w1j dj ðIÞ C w0j ð1 K dj ðIÞÞ; wj s0
>
>
>
> dj ðIÞsx
X dj ðIÞsx
X dj ðIÞsx
X
<
DðIÞ Z 0; if w1j dj ðIÞ! t ½w1j dj ðIÞ C w0j ð1 K dj ðIÞÞ; wj s0 (3)
>
> dj ðIÞsx
X dj ðIÞsx dj ðIÞsx
>
>
>
: x;
> if wj Z 0:
dj ðIÞsx

decision of the entire system since this decision becomes Let us order the units such that tj%tjC1 for 1%j%nK1
evident after the fast units have voted. This happens and define the total weight of WVS units that have the
when the total weight of units voting for the proposition decision times not greater than tm and support the
acceptance or rejection is enough to guarantee the system proposition I as Wm1
decision independently of the decisions of the units that X
m

have not voted yet. In such situations the voting process Wm1 Z w1j 1ðdj ðIÞ Z 1Þ; (4)
jZ1
can be terminated without waiting for slow units’
decisions and the WVS decision can be made in a and the total weight of WVS units that have the decision
shorter time. times not greater than tm and reject the proposition as Wm0 :
The number of combinations of unit decisions that allow X
m
the entire system decision to be obtained before the outputs Wm0 Z w1j 1ðdj ðIÞ Z 0Þ: (5)
of all of the units become available depends on the unit jZ1

weights distribution and on the threshold value. By The decision rule (3) can now be rewritten as follows:
increasing the weights of the fastest units one makes the 8
WVS more decisive in the initial stage of voting and >
> 1; if Wn1 R tðWn1 C Wn0 Þ; Wn1 C Wn0 s0
<
therefore reduces the mean system decision time by the DðIÞ Z 0; if Wn1 ! tðWn1 C Wn0 Þ; Wn1 C Wn0 s0 :
price of making it less reliable. >
>
:
In applications where the WVS should make many x; if Wn1 C Wn0 Z 0:
decisions in a limited time the expected system decision (6)
time is considered to be a measure of its performance.
Following this expression the condition that D(I)Z0 can
Since the units’ weights and the threshold affect both the
be rewritten as
WVS’s reliability and its expected decision time, the
problem of the optimal system turning can be formulated Wn1 ! tðWn1 C Wn0 Þ (7)
as follows: find the voting units’ weights and the or
threshold that maximize the system reliability R while
providing the expected decision time T not greater than ð1 K tÞWn1 K tWn0 ! 0: (8)
180 G. Levitin / Reliability Engineering and System Safety 89 (2005) 177–184

This gives one the simple way of tallying the units’ votes: that X meets some condition X2q, the coefficients of
each unit j adds value of ð1K tÞw1j to the total WVS score if the polynomial U(z) should be summed for every term
it votes for proposition acceptance, value ofKtw0j if it votes with xk2q. This can be done using the operator d as
for proposition rejection and nothing if it abstains. The follows:
proposition is rejected if the total score is negative.
X
K
Let Vi1 and Vi0 be the sum of ‘positive’ and ‘negative’ PrfX 2qg Z dðUðzÞ; qÞ Z sk 1ðxk 2qÞ: (15)
weights of units from i to n respectively kZ1

X
n X
n
In our case, the polynomial U(z) can define system unit
Vi1 Z w1j ; Vi0 Z w0j ; (9) output distribution, i.e. it can represent all the possible
jZi jZi
states of the system by relating the probabilities of each
1 0 state k to voting results corresponding to unit outputs dj(I)
(VnC1 Z VnC1 Z 0 by definition).
After the voting of the first m units the total system score (1%j%n) in that state. Each WVS state k can be
is ð1K tÞWm1 K tWm0 . The maximal possible value of the characterized by two indices: state probability sk and
score after the rest of the units add their votes is total score of unit votes at state
P k Gk.
Consider the UF UðzÞZ KkZ1 sk zGk , which relates
ð1 K tÞWm1 K tWm0 C ð1 K tÞVmC1
1
; (10) these two indices. Using this polynomial one can describe
(if all of the units from mC1 to n vote for proposition output distributions of a system consisting of an individual
acceptance) and the minimal possible value of the score is unit j as

ð1 K tÞWm1 K tWm0 K tVmC1


0
: (11) X
3
ujI ðzÞ Z sjk zGjk ; (16)
(if all of the units from mC1 to n vote for proposition kZ1
rejection). If the maximal possible score is negative
where for IZ1
ð1 K tÞWm1 K tWm0 C ð1 K tÞVmC1
1
!0 (12)
the proposition will be rejected independently of the sj1 Z qðjÞ
10 ; Gj1 Z Ktw0j ;
decisions of the units mC1,.,n. Therefore, there is no sj2 Z qðjÞ
1x ; Gj2 Z 0; (17)
need to continue the voting and the proposition can be
rejected by the system at time tm. If the minimal possible sj3 Z qðjÞ ðjÞ ðjÞ
11 Z ð1 K q10 K q1x Þ; Gj3 Z ð1 K tÞw1j ;
score is not negative
and for IZ0
ð1 K tÞWm1 K tWm0 K tVmC1
0
R0 (13)
there is no chance that the system will reject the proposition sj1 Z qðjÞ
01 ; Gj1 Z ð1 K tÞw1j ;
even if the units mC1,.,n vote for its rejection. Therefore, sj2 Z qðjÞ
0x ; Gj2 Z 0; (18)
the proposition can be accepted by the system at time tm
without waiting for decisions of units mC1,.,n. sj3 Z qðjÞ ðjÞ ðjÞ
00 Z ð1 K q01 K q0x Þ; Gj3 Z Ktw0j :
The first state corresponds to a wrong decision, the
second state corresponds to abstention and the third state
corresponds to the right decision. To obtain the UF of a
3. Evaluation of WVS reliability and expected
system containing first two voting units with their individual
decision time
UF u1I ðzÞ and u2I ðzÞthe following composition operator can
be used:
3.1. Universal generating function technique !
X3 X
3
2 1 2
UI ðzÞ Z UðuI ðzÞ; uI ðzÞÞ Z U s1i z ;
G1i
s2k z G2k
The procedure used for WVS reliability and expected
iZ1 kZ1
decision time evaluation is based on the universal moment 3 X
X 3
generating function technique. This procedure is a modifi- Z s1i s2k zG1iCG2k : ð19Þ
cation of the algorithm presented in [9]. iZ1 kZ1
The universal moment generating function of a discrete
random variable X is defined as a polynomial One can see that the resulting UF UI2 ðzÞ represents all the
possible states of the two-unit WVS. Indeed, by multiplying
X
K corresponding probabilities and by summing scores one
UðzÞ Z s k z xk ; (14) obtains polynomial with 32Z9 terms corresponding to all
kZ1
the possible states of the two-unit WVS. Each term
where the variable X has K possible values and sk is the describes individual WVS state by relating its probability
probability that X is equal to xk. To obtain the probability and the total system score.
G. Levitin / Reliability Engineering and System Safety 89 (2005) 177–184 181

Given the UF for first mK1 voting units UImK1 ðzÞ, one can and calculated as follows:
easily obtain the UF for first m voting units UIm ðzÞ as:
R Z p0 Q00 C p1 Q11 : (24)
X
Km
Gm
UIm ðzÞ Z UðUImK1 ðzÞ; um
I ðzÞÞ kz :
sm (20)
Z k

kZ1
Having the times of units’ voting and the probabilities
that the WVS decision is made after voting of m units one
Consecutively, applying Eq. (20) for mZ2, .,n one can can obtain the expected decision time as
obtain the UF for the entire WVS UIn ðzÞ. Note that while the
X
n X
n
total number of different states is 3n, many of these states ðhm 10 C h11 Þtm :
m
T Z p0 00 C h01 Þtm C p1 ðhm m
(25)
can result in the same values of score G. Therefore the total mZ1 mZ1
number of terms Kn in UIn ðzÞ can be much less then 3n
because of like terms collection. In the worst case, the
3.2. Algorithm for evaluating WVS reliability and expected
complexity of the procedure that obtains the UF UIn ðzÞ is
decision time
O(3n).
Using the d operator (15) over UIm ðzÞ one can obtain
Using the technique described in Section 3.1, one can
probabilities that the proposition I will be accepted or
obtain the system reliability and the expected decision time
rejected after voting of first m units as
by the following procedure:
hm m m 1
I0 Z dðUI ðzÞ; Gk ! ðt K 1ÞVmC1 Þ
1. Assign U00 ðzÞZ U10 ðzÞZ 1; Q00ZQ11ZTZ0.
X
Km 2. For each voting unit j define two UF uj0 ðzÞ and uj1 ðzÞ in the
Z sm m 1
k 1ðGk ! ðt K 1ÞVmC1 Þ; (21) form (16) using Eqs. (17) and (18).
kZ1 3. For mZ1,.,nC1 obtain ðtK 1ÞVm1 and tVm0 using Eq.
(which corresponds to the sum of probabilities of scores (9).
meeting condition (12) of proposition rejection) and 4. For mZ1,.,n
4.1 Obtain U0m ðzÞ and U1m ðzÞ using Eq. (20).
hm 00 ; h01 from U0 ðzÞ and h10 , h11 from U1 ðzÞ
m m 0
I1 Z dðUI ðzÞ; Gk R tVmC1 Þ 4.2 Obtain hm m m m m m

using Eqs. (21) and (22).


X
Km
4.3 Add hm m
00 to Q00, add h11 to Q11.
Z sm m 0
k 1ðGk R tVmC1 Þ: (22) 4.4 Add p0 tm ðh00 C h01 ÞC p1 tm ðhm
m m m
10 C h11 Þ to T.
kZ1
4.5 Remove the terms meeting conditions (12) and (13)
(which corresponds to the sum of probabilities of scores fromQU0m ðzÞ and U1m ðzÞ:
meeting condition (13) that proposition cannot be rejected). 5. Subtract njZ1 qðjÞ1x from Q11.
Terms meeting the conditions Gm 1
k ! ðtK 1ÞVmC1 or Gk R
m
6. Calculate WVS reliability R using Eq. (24).
0
tVmC1 correspond to the scores that guarantee that the
system can make decisions independently of decisions of
voting units mC1,.,n. Therefore, these terms can be 3.3. Example of WVS reliability and expected decision
removed from the UF UIm ðzÞ without loosing any infor- time evaluation
mation when the next UF UImC1 ðzÞ is determined.
The sum hm m
I0 C hI1 determines the probability that the Given
WVS decision is made at time tm. Summing the probabilities
hm m
I0 and hI1 for mZ1,.,n wePcan obtain the overall
n Z 3; p0 Z p1 Z 0:5; t Z 0:6;
probabilities of WVS decisions: nmZ1 hm I0 gives the
Poverall
probability that the proposition I is rejected: QI0, nmZ1 hm I1 qð1Þ ð1Þ ð1Þ ð1Þ 0 1
01 Z0:02; q10 Z0:02; q0x Zq1x Z0:01; w1 Z3; w1 Z5; t1 Z1;
gives the overall probability that the proposition I is not
rejected: QI1CQIx. (Observe that when mZn, VmC1 0
Z 0 and qð2Þ ð2Þ ð2Þ ð2Þ 0 1
01 Z0:02; q10 Z0:05; q0x Zq1x Z0:02; w2 Z4; w2 Z3; t2 Z2;
n n n
hI1 Z dðUI ðzÞ; Gk R 0Þ gives the sum of probabilities that the qð3Þ ð3Þ ð3Þ ð3Þ 0 1
01 Z0:01; q10 Z0:03; q0x Zq1x Z0:0; w3 Z3; w3 Z2; t3 Z4:
WVS accepts the proposition and that it abstains QIx).
Therefore the probability of proposition acceptance Q11
given IZ1 is obtained as follows Following step 1 of the algorithm, assign:

X
n X
n Y
n U00 ðzÞ Z U10 ðzÞ Z 1; Q00 Z Q11 Z T Z 0:
Q11 Z hm
11 K Q1x Z hm
11 K qðjÞ
1x : (23)
mZ1 mZ1 jZ1 Following step 2 of the algorithm, obtain:

Since events IZ0 and IZ1 are mutually exclusive, ð1 K tÞw11 Z 2; ð1 K tÞw12 Z 1:2; ð1 K tÞw13 Z 0:8;
the entire WVS reliability Pr{D(I)ZI} can be defined as
PrfDðIÞZ 0jI Z 0gPrfI Z 0g C PrfDðIÞZ 1jI Z 1gPrfI Z 1g Ktw01 Z K1:8; K tw02 Z K2:4; K tw03 Z K1:8:
182 G. Levitin / Reliability Engineering and System Safety 89 (2005) 177–184

The UF for the individual voting units are: U12 ðzÞ Z 10K4 ð186zK0:6 C2z0 C93z1:2 C485zK0:4 Þ:
u10 ðzÞ ¼ 10K2 ð2z2 þ z0 þ 97zK1:8 Þ; UF for subsystem consisting of three units are
u11 ðzÞ ¼ 10 ð2z
K2 K1:8 0 2
þ z þ 97z Þ; U03 ðzÞ Z UðU02 ðzÞ;u30 ðzÞÞ Z 10K6 ð192zK0:4 C2z1:2 C2z0
u20 ðzÞ ¼ 10K2 ð2z1:2 þ 2z0 þ 96zK2:4 Þ; C194zK0:6 Þð1z0:8 C99zK1:8 Þ Z10K6 ð192z0:4 C2z2:0
u21 ðzÞ ¼ 10K2 ð5zK2:4 þ 2z0 þ 93z1:2 Þ; C2z0:8 C194z0:2 C19008zK2:2 C198zK0:6
u30 ðzÞ ¼ 10K2 ð1z0:8 þ 99zK1:8 Þ; C198zK1:8 C19206zK2:4 Þ;
u31 ðzÞ ¼ 10K2 ð3zK1:8 þ 97z0:8 Þ:
U13 ðzÞ Z UðU12 ðzÞ;u31 ðzÞÞ Z 10K6 ð186zK0:6 C2z0 C93z1:2
Following step 3 of the algorithm, obtain:
C485zK0:4 Þð3zK1:8 C97z0:8 Þ
tV20 Z tðw02 C w03 Þ Z 0:6 !7 Z 4:2;
Z 10K6 ð558zK2:4 C6zK1:8 C279zK0:6 C1455zK2:2
tV30 Z tw03 Z 0:6 !3 Z 1:8; tV40 Z 0; C18042z0:2 C194z0:8 C9021z3:0 C47045z0:4 Þ:

ðt K 1ÞV21 Z ðt K 1Þðw12 C w13 Þ Z K0:4 !8 Z K3:2; In the final UF all of the terms meet either condition (12)
or (13). The terms meeting condition (12) are marked in
ðt K1ÞV31 Z ðtK1Þw13 ZK0:4!2 ZK0:8; ðt K1ÞV31 Z 0: bold italic, the terms meeting condition (13) are marked in
bold.
Following step 4 of the algorithm, obtain: h300 Ch301 Z10K6 ð192 C2 C2 C194C19008 C198
U01 ðzÞ Zu10 ðzÞ; U11 ðzÞ Zu11 ðzÞ: C198 C19206Þ Z0:039;

UF U01 ðzÞ and U11 ðzÞ do not contain terms that meet Q00 Z0:9602Ch300 Z 0:9602 C10K6 ð19008 C198 C198
conditions (12) and (13). This means that the WVS cannot
C19206Þ Z 0:99881;
make any decision based on voting of the first unit and
h100 Zh101 Zh110 Z h111 Z 0:
h310 Ch311 Z10K6 ð558 C6 C279 C1455C18042 C194
UF for subsystem consisting of two units are C9021 C47045Þ Z0:0766;
U02 ðzÞ ZUðu10 ðzÞ; u20 ðzÞÞ
Q11 Z0:9215Ch311 Z 0:9215 C10K6 ð18042 C194 C9021
Z10K4 ð4z3:2 C4z2 C192zK0:4 C2z1:2 C2z0 C47045Þ Z 0:995802:

C96zK2:4 C194zK0:6 C194zK1:8 C9312zK4:2 Þ; Since Q1x Zqð1Þ ð2Þ ð3Þ


1x q1x q1x Z0, Q11 KQ1x ZQ11 Z0:995802.
The WVS reliability is
U12 ðzÞ ZUðu11 ðzÞ; u21 ðzÞÞ R Z p0 Q00 C p1 Q11 Z 0:5 !0:99881 C 0:5 !0:995802
Z10K4 ð10zK4:2 C4zK1:8 C186zK0:6 C5zK2:4 C2z0 Z 0:997306:

C93z1:2 C485zK0:4 C194z2 C9021z3:2 Þ: The expected decision time is


T Z 0:5ð0:961 !2 C 0:039 !4Þ C 0:5ð0:9234 !2
The terms meeting condition (12) are marked in bold
italic, the terms meeting condition (13) are marked in bold. C 0:0766 !4Þ Z 2:1156:
According to steps 4.2 and 4.3 of the algorithm:
h200 Ch201 Z 10K4 ð4C4 C96C194 C9312Þ Z0:961; 4. Solving the optimization problem
Q00 Z h200 Z 10K4 ð96 C194C9312Þ Z 0:9602;
4.1. Optimization procedure
h210 Ch211 Z 10K4 ð10 C4C5 C194 C9021Þ Z 0:9234;
In order to solve the optimization problem (2) the same
Q11 Zh211 Z10K4 ð194 C9021Þ Z 0:9215: procedure is used that has been applied in [9] for solving the
After removing the marked terms the UF take the form optimization problem (1). This procedure uses a genetic
algorithm, which is based on the principle of evolutionary
U02 ðzÞ Z10K4 ð192zK0:4 C2z1:2 C2z0 C194zK0:6 Þ; search in solution space.
G. Levitin / Reliability Engineering and System Safety 89 (2005) 177–184 183

Basic notions of GA are originally inspired by Table 2


biological genetics. GA operates with ‘chromosomal’ Optimal weights for T*Z35
representation of solutions, where crossover, mutation No of 1 2 3 4 5 6
and selection procedures are applied. This representation unit j
requires the solution to be coded as a finite length string. p0Z0.7 w1j 0.018 0.240 0.564 0.300 0.388 0.476
The natural representation of a WVS weight distribution w0j 1.958 0.018 0.370 2.487 1.005 0.176
is by an 2nC1—length integer string in which the values p0Z0.5 w1j 0.017 2.367 0.497 0.017 0.635 1.475
in 2jK1 and 2j position corresponds to the weights w0j w0j 2.281 0.360 0.189 1.561 0.566 0.034
and w1j of j-th unit of the WVS and the value in position p0Z0.3 w1j 0.019 2.597 1.243 0.019 0.019 0.742
nC1 corresponds to the threshold. The unit weights are w0j 1.688 0.334 0.204 1.967 0.909 0.260
further normalized in such a way that their total weight
is always equal to some constant.
Eq. (2) the fitness function is defined as:
The detailed information on GA can be found in
Goldberg’s comprehensive book [16], and recent develop-
ments in GA theory and practice can be found in books [17, F Z R K a minðT K T*; 0Þ; (26)
18]. The basic structure of the version of GA, referred to as
GENITOR [19], is as follows. where a is a penalty coefficient. For solutions with T!T*
First, an initial population of Ns randomly constructed the fitness of the solution depends only on WVS reliability.
solutions (strings) is generated. Within this population, new The solution fitness is used to compare different solutions.
solutions are obtained during the genetic cycle by using The comparison is accomplished by a selection
crossover and mutation operators. The crossover produces a procedure that determines which solution is better: the
new solution (offspring) from a randomly selected pair of newly obtained solution or the worst solution in the
parent solutions, facilitating the inheritance of some basic population. The better solution joins the population,
properties from the parents by the offspring. The probability while the other is discarded. If the population contains
of selecting the solution as a parent is proportional to the equivalent solutions following selection, redundancies are
rank of this solution. (All the solutions in the population are eliminated and the population size decreases as a result.
ranked in order of their fitness increase). In this work, we After new solutions are produced Nrep times, new
use the so-called two point (or fragment) crossover operator randomly constructed solutions are generated to replenish
which creates the offspring string for the given pair of parent the shrunken population, and a new genetic cycle begins.
strings by copying string elements belonging to the The GA is terminated after Nc genetic cycles. The final
fragment between two randomly chosen positions from population contains the best solution achieved.
the first parent and by copying the rest of string elements
from the second parent. 4.2. Illustrative example
Each offspring solution undergoes mutation, which results
in slight changes to the offspring’s structure and maintains a A WVS consists of six voting units with voting times and
diversity of solutions. This procedure avoids premature fault probabilities presented in Table 1. The optimal voting
convergence to a local optimum and facilitates jumps in the unit weights and thresholds and the parameters of the
solution space. The positive changes in the solution code, optimal WVS obtained for T*Z35 (when p0Z0.7, p0Z0.5,
created by the mutation can be later propagated throughout p0Z0.3) are presented in Tables 2 and 3. The system
the population via crossovers. In our GA, the mutation abstention probabilities do not depend of its weights and
procedure swaps elements initially located in two randomly threshold. For any solution Q0xZ0.868!10K7, Q1xZ
chosen positions on the string. Example of crossover and 1.89!10K7.
mutation procedures can be found in [8]. It can be seen that for p0s0.5 the WVS takes advantage
Each new solution is decoded and its objective function of knowledge about statistical asymmetry of the input and
(fitness) value is estimated. In order to find the solution of provides greater reliability than in the case when p0Z0.5.
Table 1 Table 3
Parameters of voting units Parameters of optimal WVS
No of tj qðjÞ
01 qðjÞ
0x qðjÞ
10 qðjÞ
1x p0Z0.7 p0Z0.5 p0Z0.3
unit j
t 0.76 0.50 0.45
1 10 0.22 0.31 0.29 0.12 Q00 0.9798 0.9005 0.8477
2 12 0.35 0.07 0.103 0.30 Q01 0.0202 0.0995 0.1523
3 38 0.24 0.08 0.22 0.15 Q10 0.1611 0.0719 0.0283
4 48 0.10 0.05 0.2 0.01 Q11 0.8389 0.9281 0.97166
5 55 0.08 0.10 0.15 0.07 R 0.9375 0.9143 0.9345
6 70 0.08 0.01 0.10 0.05 T 34.994 34.987 34.994
184 G. Levitin / Reliability Engineering and System Safety 89 (2005) 177–184

References

[1] Ben Dov Y. Optimal reliability design of k-out-of-n systems subject to


two kinds of failure. J Oper Res Soc 1980;31:743–8.
[2] Biernat J. The effect of compensating fault models on n-tuple modular
redundant system reliability. IEEE Trans Reliab 1994;43:294–300.
[3] Pham H, Pham M. Optimal design of (k,nKkC1) systems subject to
two modes. IEEE Trans Reliab 1991;40:559–62.
[4] Pham H. Optimal system size for k-out-of-n systems with competing
failure modes. Math Comput Model J 1991;15:77–82.
[5] Satoh N, Sasaki M, Yuge T, Yanagi S. Reliability of 3-state device
systems with simultaneous failures. IEEE Trans Reliab 1993;42:
470–7.
[6] Pham H, Malon M. Optimal design of systems with competing failure
modes. IEEE Trans Reliab 1994;43:251–4.
[7] Nordmann L, Pham H. Weighted voting systems. IEEE Trans Reliab
1999;48(1):42–9.
[8] Levitin G, Lisnianski A. Reliability optimization for weighted voting
system. Reliab Eng Syst Safety 2001;71:131–8.
Fig. 2. Reliability vs. expected decision time for p0Z0.7, p0Z0.5, p0Z0.3. [9] Levitin G. Asymmetric weighted voting system. Reliab Eng Syst
Safety 2002;76:205–12.
[10] Levitin G. Maximizing survivability of vulnerable weighted voting
Observe that when p0O0.5 the WVS provides Q00 greater systems. Reliab Eng Syst Safety 2003;83:17–26.
than Q11 and, vice versa, when p0!0.5 Q00!Q11. [11] Xie M, Pham H. Modeling the reliability of threshold weighted voting
The R vs. T tradeoff curves for the WVS are presented in systems. Reliab Eng Syst Safety 2004 [in press].
Fig. 2. These curves are obtained by solving the optimiz- [12] Yacoub S. Analyzing the behavior and reliability of voting systems
comprising tri-state units using enumerated simulation. Reliab Eng
ation problem (2) for different values of time constraint T*. Syst Safety 2003;81:133–45.
[13] Chen L, Avizienis A. N-version programming: a fault tolerance
approach to reliability of software operation. Proceedings of the
5. Conclusions and further research international symposium on fault-tolerant computing, Toulouse,
France; June 1978. p. 3–9.
[14] Avizienis A, et al. The STAR (self-testing-and-repairing) computer:
It is shown that if the weighted voting system consists of
an investigation of the theory and practice of fault-tolerant computer
n units that need different time to produce their outputs, its design. IEEE Trans Comput 1971;C-20:1312–21.
decision time depends on the distribution of unit weights [15] Gogiashvili Z, Namicheishvili O, Shonia G. Optimization of weights
and on the value of the threshold and that a tradeoff exists for threshold redundancy of binary channels by the method of
between the system reliability and its rapidity. (Mahalanobis’) generalized distance. Proc MMR 2000, 2nd Int Conf
An algorithm that finds the system parameters maximiz- Math Methods Reliab 2000;463–6.
[16] Goldberg D. Genetic algorithms in search, optimization and machine
ing its reliability under constraint imposed on the expected learning. Reading, MA: Addison Wesley; 1989.
system decision time is suggested. The algorithm is based [17] Back T. Evolutionary algorithms in theory and practice. Evolution
on the universal generating function technique. The strategies. Evolutionary programming. Genetic algorithms. New
computational complexity of the algorithm is in the worst York: Oxford University Press; 1996.
case O(3n). [18] Gen M, Cheng R. Genetic algorithms and engineering design. New
York: Wiley; 1997.
Further research can be devoted to analysis of weighted
[19] Whitley D. Algorithm and selective pressure: why rank-based
voting systems composed from voting units with variable allocation of reproductive trials is best. In: Schaffer D, editor.
decision times, systems with dependent voting units, Proceedings of the third international conference on genetic
systems with weights adaptation, etc. algorithms. Los Altos, CA: Morgan Kaufmann; 1989. p. 116–21.

You might also like